Research map / Vision-language and multimodal / Human and embodied modeling
SAEs-BrainMap: Unveiling the Emergence of Specialized Concepts in Deep Models via Brain Alignment
Ziming Mao, Jia Xu, Wenxuan Pan, Mufan Xue, Yaochu Jin, Guoyuan Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CLS-trained SAE features instead of raw neurons (ablation) | ViT-B/16 CLIP · NSD S5 | max cosine similarity: 0.5875 -> 0.7247 | Improved |
| use CLS-trained SAE features instead of raw neurons (ablation) | ViT-B/16 CLIP · NSD S5 | RSA score: 0.511 -> 0.580 | Improved |
| use patch-trained SAE features instead of raw neurons (ablation) | DINOv2 · NSD S5 | max cosine similarity: 0.6961 -> 0.6517 | Hurt |
| use patch-trained SAE features instead of raw neurons (ablation) | ImageNet ViT · NSD S5 | RSA score: 0.596 -> 0.656 | No clear effect |
| select features with brain fMRI guidance instead of CLIP text matching (CLIP-Dissect, CLIP-SAEs) (ablation) | ViT-B/16 CLIP · face-selective features | layer of emergence and semantic coherence | Improved |
| train SAEs on all layers to trace concept emergence | ViT-B/16 CLIP · Face concept | emergence layer: emerges at Layer 2 | Improved |
Closest papers by meaning
- Sparse components distinguish visual pathways & their alignment to neural networks (ICLR 2025 · similarity 0.89)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.88)
- A Convolutional Neural Network Interpretable Framework for Human Ventral Visual Pathway Representation (AAAI 2024 · similarity 0.88)
- Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts (NeurIPS 2025 · similarity 0.87)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.87)
- Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
